24.02.2013 Views

Optimality

Optimality

Optimality

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

optimal<br />

1<br />

2<br />

3<br />

4<br />

leaf sets 5<br />

6<br />

(leaf set size)<br />

Optimal sampling strategies 279<br />

1<br />

2<br />

3<br />

optimal 4<br />

leaf sets 5<br />

6<br />

(leaf set size)<br />

unbalanced<br />

variance of<br />

innovations<br />

(a) Scale-invariant tree (b) Tree with unbalanced variance<br />

1<br />

2<br />

optimal 3<br />

leaf sets<br />

4<br />

5<br />

6<br />

(leaf set size)<br />

(c) Tree with missing leaves<br />

Fig 4. Optimal leaf node sets for three different independent innovations trees: (a) scale-invariant<br />

tree, (b) symmetric tree with unbalanced variance of innovations at scale 1, and (c) tree with<br />

missing leaves at the finest scale. Observe that the uniform leaf node sets are optimal in (a) as<br />

expected. In (b), however, the nodes on the left half of the tree are more preferable to those on<br />

the right. In (c) the solution is similar to (a) for optimal sets of size n = 5 or lower but changes<br />

for n = 6 due to the missing nodes.<br />

We consider a symmetric tree in Fig. 4(b), that is a tree in which all nodes have<br />

the same number of children (excepting leaf nodes). All parameters are constant<br />

within each scale except for the variance of the innovations Wγ at scale 1. The<br />

variance of the innovation on the right side is five times larger than the variance<br />

of the innovation on the left. Observe that leaves on the left of the tree are now<br />

preferable to those on the right and hence dominate the optimal sets. Comparing<br />

this result to Fig. 4(a) we see that the optimal sets are dependent on the correlation<br />

structure of the tree.<br />

In Fig. 4(c) we consider the same tree as in Fig. 4(a) with two leaf nodes missing.<br />

These two leaves do not belong to the optimal leaf sets of size n = 1 to n = 5 in<br />

Fig. 4(a) but are elements of the optimal set for n = 6. As a result the optimal sets<br />

of size 1 to 5 in Fig. 4(c) are identical to those in Fig. 4(a) whereas that for n = 6<br />

differs. This result suggests that the optimal sets depend on the tree topology.<br />

Our results have important implications for applications because situations arise<br />

where we must model physical processes using trees with different correlation structures<br />

and topologies. For example, if the process to be measured is non-stationary<br />

over space then the multiscale tree may be unbalanced as in Fig. 4(b). In some<br />

applications it may not be possible to sample at certain locations due to physical<br />

constraints. We would thus have to exclude certain leaf nodes in our analysis as in<br />

Fig. 4(c).<br />

The above experiments with tree-depth D = 3 are “toy-examples” to illustrate<br />

key concepts. In practice, the water-filling algorithm can solve much larger real-

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!